XNews

Databricks Releases MemCurrent Persistent Memory Architecture for Production Agents

Databricks Releases MemCurrent Persistent Memory Architecture for Production Agents

Databricks released MemCurrent, a persistent memory architecture for production agents that combines episodic and derived memory across plugins, on-demand tools and an offline dreaming pipeline. On an internal benchmark built from real enterprise data, MemCurrent triples the agent pass rate while using 31% fewer tokens and running 26% faster. It is built into Agent Bricks' Memory offering and already powers memory in Genie.

Event version 1.

Key insights

Two memory types

MemCurrent combines episodic memory, raw session trajectories recorded append-only by a plugin, with derived memory, distilled experience stored as hierarchical files topped by an auto-generated memory.md index.

Three access patterns

The architecture uses plugins that log and load context automatically at fixed session points, tools the agent calls on demand, and an offline dreaming pipeline that consolidates memory in the background.

Benchmark results

On an internal benchmark built from over 12,000 real Genie sessions and filtered to 114 test cases, pass rate rose from 22.3% without memory to 68.4%, against an ideal ceiling of 78.6%, while running 1.35 times faster and 1.44 times cheaper.

Public benchmark comparison

On LoCoMo, LongMemEval and ConvoMem, MemCurrent averaged 88.1%, comparable to full-context 87.8%, while using about 12 times fewer input tokens per question, 8.4K versus 103K.

Ablation findings

Episodic memory alone reached 35.8%, derived memory with an index reached 66.1%, and combining both reached 68.4%; offline dreaming alone reached 63.5%, online tools alone 53.9%, and both combined 66.1%.

Source posts